From detection to forecasting: Utilizing time-series foundation models to anticipate defects in metal additive manufacturing

Abstract

Metal additive manufacturing (MAM) offers material efficiency, customization, rapid prototyping, and complex geometries, but challenges in reliability and consistent quality hinder its widespread industrial adoption, necessitating robust quality assurance mechanisms. Traditional machine learning (ML)-based in situ monitoring (ISM) essentially detects existing defects but is hindered by inadequate generalization ability and low training efficiency. Motivated by recent advancements in time-series foundation models for general time-series analysis, this study introduces defect forecasting in MAM using time-series foundation models to address these issues. Three representative models (GPT4TS, TimeLLM, UniTS) are employed across two ISM datasets with different sensor data and processes. Through fine-tuning, we demonstrate strong generalizability and competitive performance with minimal training and compact input data. Compared to state-of-the-art models, our approach offers efficient training and enhanced generalizability. Our contributions include leveraging time-series foundation models in MAM ISM for defect forecasting, addressing key challenges in traditional ML-based ISM, and demonstrating efficient results across tasks and datasets. Thus, this work advances ML applications in ISM by shifting from defect detection to forecasting, implying the possibility of proactive defect prevention in MAM.